Open exploration into continual learning for agents.

Every agent deployment we've seen follows the same arc.

The agent ships. It fails. The team patches the prompt, adjusts the retrieval, adds a guardrail. It fails again, differently, but just as predictably. The failures get logged. Nobody uses the logs. Nothing actually learns.

At small scale, this is a papercut. At the scale enterprises are moving toward, it's the reason AI initiatives stall.

The instinct is to blame the model. Or the prompts. Or the infrastructure. But we think the problem is more fundamental: agents have no mechanism to accumulate experience. They can't know what went wrong last time, in a situation like this one. Every deployment starts from zero.

We're exploring what it would take to fix that, building infrastructure that gives agents structured memory from their own production history, so that every failure makes the next decision better.

We don't think this is a fully solved problem. We think it's one of the more important unsolved ones. And we're doing all of it in the open.

If that's the kind of problem you want to work on, as a builder or a researcher who believes experience should compound, we'd like to talk.